Two out of three AI answers are good enough to send right away.
A language model can write a friendly answer in seconds. The hard part is making sure the answer is correct: with the right product knowledge, support rules, current customer information and the status of technical issues. Robo Helma brings those sources together and drafts an answer that a support agent reviews before it goes to the customer.
The case in brief
67% sent without edits
18%
15%
Skills, vector search and live context
Where was the real problem?
Automating support looks like a prompt problem at first sight. In practice, the difficulty lies elsewhere. A good answer requires different kinds of knowledge, each with its own source, freshness and reliability.
See the approachProduct knowledge changes constantly
FAQs, manuals, policies, known issues and other documentation must be available to the AI without stuffing everything into one huge prompt or hard-coding knowledge by hand.
Not all knowledge is documentation
Tone of voice, support policy, escalation rules and ways of working are instructions. Customer data and current incidents are live data. You cannot reliably treat these different kinds of context as if they were all the same.
A convincing answer can still be wrong
A model can neatly make up something that sounds plausible. For support, that is not enough. The AI must know which sources are leading, when information is missing and when a person must review or take over.
How did we approach it?
- 01
Split knowledge by function
We treat behaviour and support rules as OpenAI Skills, product documentation as searchable knowledge in a vector store and current customer or system information as external context. This gives every kind of information a clear place.
- 02
Retrieve context before answering
For each support question, Helma first gathers the relevant context. Product knowledge comes via File Search, customer information can be retrieved through an external MCP connection and current incident information comes from a separate controlled source.
- 03
Connect engineering and support
A support ticket can be linked to an existing GitHub issue. Helma follows the technical status read-only. When engineering closes or reopens an issue, the new status can be included in a customer-safe draft answer without exposing internal development information.
- 04
People keep the final say
Helma drafts an answer for human review. The support agent can send it directly or edit it. Both the original AI answer and the answer that was eventually sent are stored, so we can measure how useful the AI really is.
What did we build?
Robo Helma is a complete support workflow in which AI, knowledge, integrations, ticketing and feedback come together, not a standalone chatbot next to an existing helpdesk.
A configurable knowledge layer
Every brand can use its own OpenAI Skills and its own vector store. Skills contain, for example, support behaviour, tone of voice and escalation policy. The vector store holds product knowledge such as FAQs, policies, known issues and documentation.
Context from real systems
Helma can combine product knowledge with customer-specific information, current incidents, previous customer interactions and linked engineering context. That way the model does not have to make up critical facts itself.
A complete support environment
Around the AI we built a multi-tenant inbox, ticket statuses, internal notes, attachments, reporting, feedback tracking and integrations with, among others, email, Facebook Messenger, app store reviews and GitHub.
Reliable AI takes more than good prompts
Customers can put anything in a message and external sources can contain outdated or unreliable information. That is why we do not automatically treat that input as an instruction for the model.
Prompt injection is checked explicitly
Helma checks customer messages for attempts to manipulate instructions, safety rules, tools or hidden prompts.
Every source has its own role
Product documentation determines product facts, customer data determines customer-specific information and GitHub only provides technical status. One source cannot simply override another.
No automatic promises from engineering data
A closed GitHub issue, for example, does not automatically mean that a fix has been deployed or is available. That boundary is explicitly built into the AI workflow.
What does it deliver?
67% can go straight to the customer
Currently, 67% of the support answers generated by Robo Helma are sent without edits. The support agent only needs to review and send the answer.
Another 18% only needs an edit
For 18%, the generated answer is largely usable, but a support agent adjusts something before it is sent.
The remaining 15% shows where we can improve further
About 15% of the answers are fully rewritten. Because we store the original AI draft and the final answer, we can see exactly where Helma still falls short and keep improving on that.
Proof
The reliability comes from the engineering around the model. Robo Helma stores the context used for every AI run and measures what support agents eventually do with the generated answer.
View cases67% direct AI acceptance
Two out of three generated answers are currently sent without substantive edits. Another 18% is only slightly adjusted.
Source information is visible
In the AI analysis, a support agent can see which knowledge references were used via File Search, including the files used and available relevance scores.
Improvement is part of the system
Reporting compares the generated draft with the answer that was actually sent and classifies it as accepted, edited or rejected. That way we can measure whether the AI is getting better and where improvement is needed.
Related AI challenges
Have an AI agent built
You know which task or workflow the agent has to perform. Now it needs to be designed well, connected and made reliable enough for real use.
View use caseHave an internal AI assistant developed
You want to give employees faster access to internal knowledge, documents and systems through an assistant that fits your organisation.
View use caseMake your own documents, data and knowledge usable for AI
The valuable information is already there. Now AI has to be able to work with it reliably, securely and practically.
View use caseAdd AI to existing software, an app or SaaS
The product already exists. Now you want to add AI functionality without losing sight of the existing architecture, user experience and reliability.
View use caseWe're missing tests, evals or monitoring
The AI functionality runs, but nobody can objectively see whether changes make it better, worse or quietly break something.
View use caseBuild & Integrate
Do you have an AI challenge of your own where multiple knowledge sources, existing systems and reliable AI need to come together? Then we design and build the workflow and integrations around your situation.
View offerFixed price after scoping
- Result
- Production-ready AI integrated into your existing environment
- Duration
- Depends on scope
- Next step
- First we determine which knowledge, systems and controls are really needed